See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/337991478

Modeling and Forecating the Markets Volatility and VAR Dynamics of

Commodity

Article · January 2017

CITATION READS

1 68

2 authors:

Pınar Kaya Bulent Guloglu

Marmara University Istanbul Technical University
8 PUBLICATIONS   28 CITATIONS    21 PUBLICATIONS   243 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Financial Contagion and flight to quality View project

All content following this page was uploaded by Pınar Kaya on 22 May 2020.

The user has requested enhancement of the downloaded file.

 Modeling and Forecasting the Markets
Volatility and VaR Dynamics of Commodity
Pınar KAYA*
Bülent GÜLOĞLU**

Abstract
The purpose of this paper is to model and forecast the risk of six commodities na-
mely, crude oil, copper, gold, silver, palladium, and platinum during the period from
02/01/2002 to 29/04/2016 using volatility, value at risk and expected shortfall as risk
measures. After showing that squared returns of all six commodities have a significant
long memory, the volatility, the value at risk and expected shortfall based on fractional
GARCH models are estimated and forecasted. Both forecast performance of volatility
models and backtest for value at risk indicate that in many cases FIAPARCH model out-
performs the other GARCH models. Then volatility, value at risk and expected shortfall
estimates based on FIAPARCH model show that the volatility and market risk of oil is
much higher than the other commodities. This casts doubt on the use of oil as a hedging
tool.
Keywords: Volatility Modelling, Commodity Markets, VaR Forecasting, Expected
Shortfall
JEL Classification: C22, C53, C58, G17, Q02

Özet - Emtia Piyasalarının Oynaklık ve Riske Maruz Değer

Dinamiklerinin Modellenmesi ve Öngörüsü

Bu çalışmanın amacı, ham petrol, bakır, altın, gümüş, paladyum ve platinden oluşan
altı temel emtiaya ait zaman serisinin 02/01/2002 - 29/04/2016 arasını kapsayan
dönemde, oynaklık, riske maruz değer ve beklenen açık risk ölçümlerini kullanarak, em-
tia piyasalarının riskini modellemek ve öngörmektir. Bu altı emtianın getiri karelerinin
önemli ölçüde uzun hafıza özelliğine sahip olduğu gösterildikten sonra oynaklık, riske
maruz değer ve beklenen açık, kesirli bütünleşik GARCH modelleri kullanılarak tahmin
edilmiş ve öngörülmüştür. Hem oynaklık modellerinin öngörü performansı hem de riske
maruz değer için yapılan geri testler birçok durumda FIAPARCH modelinin diğer GARCH
modellerinden bariz biçimde üstün olduğunu göstermektedir. FIAPARCH modelini kul-
lanarak yapılan oynaklık, riske maruz değer ve beklenen açık tahmin sonuçları petrolün
diğer emtialardan daha riskli olduğunu göstermekte ve petrolün riskten korunma aracı
olarak kullanılmasını sorgulanır hale getirmektedir.
Anahtar Kelimeler: Oynaklık Öngörüsü, Emtia Piyasaları, VaR Öngörüsü, Beklenen
Açık
JEL Sınıflandırması: C22, C53, C58, G17, Q02

* Dr., Marmara University

**Prof. Dr., Istanbul Technical University

Journal of BRSA Banking and Financial Markets

Volume: 11, Number: 1, 2017 9
 1. Introduction
In the light of recent and still growing literature of long memory property, this
paper investigates the conditional volatility of six major commodities, namely, crude
oil, copper, gold, silver, palladium, and platinum. Various forms of GARCH-class mo-
dels (linear and non-linear) allowing for asymmetry and long memory characteristic
of returns are used for this purpose.

Over recent years, we have witnessed an acceleration of financial integration

across stock markets and other assets. This in turn has weakened diversification
opportunities across financial markets and brought about the financialization of
commodity markets. Especially gold and oil have become important hedging tools
for investors seeking to diversify their risk (Kang et al. 2016). Moreover, it is well
known that gold is considered as a safe asset worldwide. In risk management, port-
folio allocation, and hedging, precious metals have been widely used as assets. The
appropriate risk measurement of an asset is therefore crucial both for investors and
fund managers. Volatility and value at risk of an asset are important indicators for
investors and would play an important role for short and long run trading decisions
along with the attitude of investors towards risk. Over recent years in addition to
Value at Risk, Expected Shortfall has also been widely used as a risk measurement
in the literature. Since expected shortfall (henceforth ESF) is the expected value of
the losses conditional on the loss being larger than the VaR (Scaillet, 2004), it is also
known as conditional value at risk. As pointed out by Tsay (2010), VaR estimates
the potential financial loss. However when an extreme event happens it could un-
derestimate the actual loss. Thus for appropriate measurement of risk, ESF besides
the VaR should be used.

There are several ways to measure market risk in the literature. In this study all
risk measures (volatility, VaR, ESF) have been based on the same GARCH models to
avoid inconsistency between different risk measurements and make them compa-
rable.

In this framework, forecasting of volatility becomes essential for appropriate me-

asurement of risk. This leads us to appropriate modeling of volatility to measure and
forecast the volatility. To address this issue we employ various linear and non-linear
GARCH models allowing for asymmetry and long memory and compare their fore-
casting performance.

This study contributes to the literature in several ways. First, we employ three dif-

10 Pınar KAYA, Bülent GÜLOĞLU

ferent risk measures for commodities. More precisely we use dynamic volatility, VaR
and ESF. All of them are coherent as they are based on the same GARCH models.
Second, forecast performance of GARCH models and backtests for VaR are carri-
ed out. Third unlike the previous studies, which use fractionally integrated GARCH
models without testing for long memory for commodities, in this study we have
employed various tests for long memory in returns and squared returns. Fourth the
results have also practical implications for investors in the sense that using appropri-
ate GARCH models they could accurately measure and forecast the future volatility.
Then the decision to invest on the commodity market could be taken. In addition,
comparing with the risk of other assets (such as equity) commodities could be used
as hedging instruments.

This paper is structured as follows. After the introduction, the second section is
devoted to the literature. The third section presents data and descriptive statistics.
The methodology used in the paper is illustrated in section four. The fifth section
presents empirical findings. The last section concludes the paper.

2. Literature
Recent empirical studies on commodity markets volatilities focus on the long
memory behavior of data (Aloui and Mabrouk, 2010). In this respect, several studies
have been carried out allowing for the long memory to model both price variations
and price volatilities.

Firstly, Mandelbrot (1971) analyzes long memory in financial markets and sug-
gests using Hurst’s ‘rescaled range’ statistic to test for long memory process in fi-
nancial return series. He argues that arbitrage opportunities might exist since in the
presence of long memory shocks cannot be absorbed quickly. As shown by Yajima
(1985), if security prices expose long memory, then the standard testing procedures
of asset pricing and the martingale models will be inappropriate. In addition, Lo
(1991) states that both the capital asset pricing model and the arbitrage pricing
theory based on standard testing procedures may not be valid if the security returns
display long memory behavior.

Long memory models are known with their autocovariances falling into decay
slowly. According to Ding et al. (1993) fully decaying of a shock can last for long.
Long memory property can be clearly explained as a significant correlation between
distant observations of time series. This means that a shock cannot be eliminated
immediately by the market. Thus, the distinction between I(0) stationary and I(1)

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 11
 nonstationary processes is prominently too restrictive (Baillie et al., 1996). While the
propagation of shocks in a stationary process arises at an exponential rate of decay,
for a nonstationary process shocks exhibit infinite persistence.

In order to bridge the gap between short and complete persistence while mo-
deling the conditional mean, fractionally integrated autoregressive moving avera-
ge (ARFIMA) specifications have been suggested. In this class of the model, the
short-run conduct of time-series is represented by the ARMA parameters, while the
fractional differencing parameter allows for the long-run dependence (Conrad and
Haag, 2006).

According to fractional integration theory, the order of integration namely the

fractional difference parameter is a fractional value (Baillie, 1996). Fractionally integ-
rated processes are different from both stationary and unit root processes with their
persistence and mean reverting features. Especially, the long memory parameter is
given by d(0, 0.5). When d > 0.5 the time series is considered as nonstationary and
when d(-0.5, 0) the series is considered as antipersistent (Kumar, 2014).

Long-memory property can occur in volatility of financial returns as well, which

points out the wide-distant correlation of time-varying volatility elements. Robinson
(1991) comes up with the linear autoregressive conditional heteroskedastic model
(LARCH) which permits long memory in the conditional variance. Later extensions of
generalized ARCH (GARCH)-type models taking into account long-memory behavior
have been proposed by many researchers. Baillie, Bollerslev, and Mikkelsen (1996)
develop the fractionally integrated GARCH (FIGARCH) model which are more con-
venient for this type of data in various empirical analyses (Bollerslev and Mikkelsen,
1996; Beine and Laurent 2003; Conrad and Karanasos, 2005a, b).

Following the studies by Granger (1980), Granger and Joyeux (1980) and Hos-
king (1981), extended use of long memory model is observed in empirical studies.
Two strands of literature in the field of long memory have received a great amount
of attention. One strand investigates modeling the volatility of commodities. Based
on standard volatility models, previous studies on the volatility of commodity focus
on single commodity’s dynamic volatility characteristics or volatility spillovers across
various commodities. More recent studies, however, extend the existing literature
by incorporating a variety of volatility behavior of several commodities (Arouri et al.,
2012a, b; Wei et al., 2010).

12 Pınar KAYA, Bülent GÜLOĞLU

 The other strand of literature focuses on the relationship between commodities
and stock markets. Kang et al.(2017) investigate spillover effects among six com-
modity futures markets (gold, silver, crude oil, corn, wheat, and rice) for the period
from January 4, 2002 to July 28, 2016. They find out bidirectional return and volatil-
ity spillovers between commodity markets. The findings of the study show that both
gold and silver are information transmitters to other commodity futures markets.
With respect to crude oil, corn, wheat, and rice, the results indicate that those com-
modities are receivers of spillovers during the financial crisis period in 2017.

Jain and Biswal(2016) examine the relationships between gold, oil USD-INR(Indian
Rupee) exchange rate and SENSEX. Using the daily data spanning the period 2006-
2015 and employing the DCC GARCH model and linear and non-linear causality
tests, they show that a decrease in gold and oil prices causes the exchange rate to
depreciate. The results of the study also show that gold becomes an investment
asset class among the investors. Besides, the authors argue that gold and oil can
be used as hedging tools against to volatility in SENSEX and fluctuations in the ex-
change rate.

Creti et al. (2013) analyze the links between returns for 25 commodities and
stocks over the period from January 2001 to November 2011. Using the dynamic
conditional correlation (DCC) GARCH model, they show that the correlations be-
tween commodity and stock markets are highly volatile especially since the 2007–
2008 financial crisis and vary through time.

In a recent study by Dahl and Iglesias (2009), an alternative functional relation-

ships (from GARCH (1, 1) to GARCH (1, 1)-AR (m)) are used to model the spot price
risk and spot prices. Based on the classical rational expectations and ARCH-M model
proposed by Engle et al. (1987), they investigate empirically the rational expecta-
tions model of Muth’s (1961). In their paper, they conclude that lagged conditional
variance should enter into the mean equation.

Kang and Yoon (2013) explore the effectiveness of a volatility model for three
crude oil markets, namely, Brent, Dubai, and West Texas Intermediate (WTI) regar-
ding persistence and long memory. Relying on the conditional volatility models they
found that the CGARCH and FIGARCH models are better equipped to capture persis-
tence and provide superior performance in terms of out-of-sample volatility forecasts
in comparison with the GARCH and IGARCH models.

In their paper, Thuraisamy et al. (2013) investigate spillover effects between the

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 13
 volatility of Asian equity market and that of the crude oil and gold futures. Their
results show that volatility shocks in fully-fledged equity markets overflow into the
crude oil and gold futures markets, while fledgling markets are in a tendency to spill
over from commodity futures to equity markets. They also provide strong evidence
of increased bi-directional volatility transmission during the 2008 financial crisis peri-
od. As for equity market volatility, not only volatility shocks from the crude oil futu-
res market are important but also the volatility shocks from the gold futures market.

In a recent study, Vivian and Wohar (2012), investigate whether there are struc-
tural breaks in commodity spot return volatility using an iterative cumulative sum of
squares procedure and the GARCH (1,1) model for each regime. Their findings pro-
vide very limited evidence of commodity volatility breaks during the recent financial
crisis compared to the 1985–2010 sample period as a whole.

In their paper, Aloui and Mabrouk (2010) appraise the value-at-risk (VaR) for
some major crude oil and gas commodities for both short and long trading po-
sitions. They calculate the VaR for various ARCH/GARCH-type models, namely FI-
GARCH, FIAPARCH and HYGARCH. Their findings show that considering long-range
memory, fat-tails, and asymmetry performs better in predicting a one-day-ahead
VaR for both short and long trading positions.

Wang et al. (2010) use two non-parametric methods, detrended fluctuation

analysis (DFA) and rescaled range analysis (R/S). They check the long memory pro-
perties of conditional volatility series obtained from GARCH-class models against
actual volatility series for WTI crude oil returns. As they are interested in the long
memory of volatility, GARCH models are appropriate for the time scale larger than
a year.

Choi and Hammoudeh (2009) conclude that forecasting the commodity volatility
relying on long memory univariate GARCH models is more accurate than the stan-
dard GARCH models for oil and refined products markets.

Other papers allowing for structural breaks and long memory properties in return
and volatility of commodity markets show that the volatility of precious metals is bet-
ter represented by long memory than by structural breaks (Arouri et al., 2012a,b).

The studies, which concentrate on commodity volatility, have been becoming

crucial because of the increasing importance of commodities in the financial sector
(Regnier, 2007; Chkili et al. 2014). Not only financial sector but also real sector,
and economic growth are significantly affected by prices of commodities. Another

14 Pınar KAYA, Bülent GÜLOĞLU

reason for their popularities are their empirical features such as their pronounced
excess kurtosis or fat tailedness, asymmetry, and structural breaks that influence the
model fitting, these series need to be estimated with varied volatility models (Aloui
and Mabrouk, 2010; Cheng and Hung, 2011; Cheong, 2009; Hung et al., 2008).
Existing researches on volatility forecasting for commodity markets are limited since
they have concentrated more on forecasting conditional return than conditional
volatility.

3. Data
In this study, daily data of six major commodity spot prices; gold, silver, copper,
crude oil, palladium, and platinum are used. Prices of two main metals gold and sil-
ver are determined by Handy & Harman (H & H) which is operated as dealers in silver
and gold and these base prices are taken for transactions worldwide. These precious
metals are measured in US dollars per troy ounce. The other precious metals cop-
per and platinum are traded on the London Metal Exchange (LME) as the second
largest traded contract. West Texas Intermediate (WTI) crude oil produced in the
US serves as a reference price in the oil market. The data set is extracted from the
Thomson and Reuters DataStream database, and the whole sample period spans
from 02/01/2002 to 29/04/2016. All estimations are based on the daily period
from January 2, 2002 to February 26, 2016. The out-of-sample forecast performan-
ces of the competing long memory-based GARCH models are based on the period
from February 29, 2016 to April 29, 2016. Daily return series are calculated as log
differences in price levels as follows:

100 × ln Pi ,t − ln Pi ,t −1 
rt = (1)

The return series of gold, copper, silver, crude oil, palladium, and platinum are
plotted in Figure 1. As illustrated in Table A1, all returns series are stationary accor-
ding to ADF and KPSS test statistics.

Descriptive statistics for the commodity return series are shown in Table 1. As
seen, all series have high excess kurtosis and negative skewness. In addition, the
distribution of the series seems to be leptokurtic. Furthermore, Ljung-Box Q (20)
statistic for returns are significant except for silver and it is significant for squared
returns of all commodities. As seen, regardless of the lag chosen, all ARCH test
statistics are highly significant. As a result, while modeling the returns, conditional
heteroscedasticity and serial correlations are taken into account.

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 15
 Figure 1: Time series plots of spot price series

16 Pınar KAYA, Bülent GÜLOĞLU

Figure 2: Time series plots of spot return series

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 17
 Table 1: Descriptive Statistics

  GOLD COPPER SILVER CRUDE_OIL PALLADIUM PLATINIUM

 Mean  0.040513  0.031673  0.031497  0.013397 0.0031338 0.017
 Maximum  6.855528  11.72590  13.66480  16.41370 12.846 8.427
 Minimum -9.596 -10.358 -12.982 -12.827 -17.859 -17.2
 Std. Dev.  1.170235  1.786288  2.023908  2.380344 2.1148 1.42
 Skewness -0.407524 -0.134399 -0.565387 -0.020894 -0.42313 -0.82
 Kurtosis  8.248861  7.025565  7.966751  7.369698 5.5092 10.21
 Jarque-Bera  4342.735  2505.367  3993.717  2939.193 4781.9 16481
ARCH 1-2 31.544[0.000] 211.86[0.000] 76.64[0.000] 145.9[0.000] 61.7[0.000] 126.7[0.000]
ARCH 1-5 41.230[0.000] 123.61[0.000] 39.59[0.000] 128.06[0.000] 38.6[0.000] 61.7[0.000]
ARCH 1-10 33.808[0.000] 74.510[0.000] 24.39[0.000] 69.27[0.000] 23.9[0.000] 34.3[0.000]
Q(20) 36.46[0.013] 52.95[0.000] 22.09[0.335] 57.13[0.000] 36.03[0.015] 24.7[0.209]
Q2(20) 1150.08[0.000] 3197.8[0.000] 635.34[0.000] 2803.92[0.000] 646.8[0.000] 964.5[0.000]
Observations  3693  3693  3693  3693  3693 3693

Figure 3: Density plots of return series

4. Methodology
In this section, the tests of long memory as well as various forms of GARCH-type
models allowing for asymmetry and long memory characteristic of commodity price
are used to model volatility. Then based on the GARCH models value at risk and
expected shortfall is calculated.
4.1. Long Memory Tests
The presence of long memory in the data implies the persistence of observed
autocorrelations. Long memory in volatility is an important phenomenon since it is
characterized by a slowly decaying autocovariance function. When long memory in
time series exists then various forms of models can be employed taking into account
intermediate degrees of volatility persistence.

18 Pınar KAYA, Bülent GÜLOĞLU

 In this paper, the Hurst-Mandelbrot Rescaled Range (R/S) statistics, Lo (1991)
Rescaled Range R/S, Geweke and Porter-Hudak (1983) (GPH), and the Robinson and
Hendry (1999) Gaussian Semiparametric (GSP) test statistics are used in order to test
for long memory components in the returns. In addition, considering long memory
in the volatility process, these tests are applied to commodities’ squared returns,
which are widely regarded as a proxy of conditional volatility (Choi & Hammoudeh,
2009; Lobato & Savin, 1998). These tests have been extensively used in the related
literature.

4.1.1. Rescaled Range (R/S) statistics

Rescaled range statistic R/S is one of the oldest and famous tests which was int-
roduced by Mandelbrot and Wallis (1969) and Hurst (1951) to examine the presen-
ce of long-term memory in time series. Mandelbrot (1971) suggests that R/S analy-
sis can be used in economic and financial investigations. Basically, Rescaled range
statistic R/S is the range of partial sums of deviations of a time series from its mean
scaled with its standard deviation. Hence, Hurst exponent H stands for the scaling
behavior of the range of cumulative departures of a time series from its mean. This
statistic is robust to data non-normality, but in the presence of autocorrelation, the
coefficients may be biased. Therefore, Mandelbrot’s null hypothesis is “there is no
long-term dependence” under the assumption of no autocorrelation.

Consequently, Lo (1991) developed a modified rescaled range, which adjusts

for possible short term dependence by applying the Newey West heteroscedasticity
and autocorrelation consistent estimator instead of the sample standard deviation.
Hence, Lo’s null hypothesis is “there is no long-term dependence”.

Geweke and Porter-Hudak (1983) propose a semi-nonparametric approach to

test for long memory with regard to a fractionally integrated process. Furthermore,
Geweke and Porter-Hudak (1983) use Fourier transformation and spectral density
into the test equation. GPH test for the null hypothesis is “there is no long memory
(d=0)”.

Gaussian Semiparametric estimation model developed by Robinson and Henry

(1999) depends on low-frequency periodogram estimates and the specification of
the shape of the spectral density of the time series. Robinson and Henry (1999)
demonstrate that the Gaussian semiparametric estimator is asymptotically normally
distributed and it is robust to conditional heteroskedasticity including the long-range
dependence.

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 19
 4.2. Results of Long Memory Tests
As can be seen from the Tables 2-7, the long memory property in return and squ-
ared returns are analyzed using test statistics of Hurst-Mandelbrot R/S, Lo R/S, GPH,
and GSP. The results show that these test statistics do not reject the null hypothesis
of no long-range dependence. So, it seems that the commodity return series does
not exhibit a long memory effect.

On the other hand, for the squared returns, tests lend a support to long memory
effect at the 1% level. Since d parameter is significantly lies into interval (0, 0.5) one
may conclude that the squared return series follow the long memory process.

Other studies dealing with commodity markets also attain almost identical results
(Aloui and Mabrouk, 2010; Cheong, 2009; Choi and Hammoudeh, 2009; Moham-
madi and Su, 2010; Wei et al., 2010).

To sum up, according to the long memory test results, the most appropriate
method to model and forecast of commodity volatility is the GARCH-class models
allowing for long memory property.

Table 2: Long Memory Test for Gold return and gold squared return

 RETURN Hurst-mandelbrot R/S Lo R/S GPH GSP

d parameter - - 0.000377 (0.0163172) -0.0118 (0.0116342)
Test Statistics 1.46717 1.47824
Critical values Probability Probability
90% [0.861, 1.747] [0.9815] [0.3088]
95% [0.809, 1.862]
99% [0.721, 2.098]
SQUARED RETURN Hurst-mandelbrot R/S Lo R/S GPH GSP
0.0702662
d parameter - - 0.125672 (0.0116342)
(0.0163172)
Test Statistics 4.51543 4.29048
Critical values Probability Probability
90% [0.861, 1.747] [0.0000] [0.0000]
95% [0.809, 1.862]
99% [0.721, 2.098]

20 Pınar KAYA, Bülent GÜLOĞLU

Table 3: Long Memory Test for Copper return and Copper squared return

RETURN Hurst-mandelbrot R/S Lo R/S GPH GSP

d parameter - - -0.0205351 (0.0163172) -0.022516 (0.0116342)
Test Statistics 1.39618 1.44774
Critical values Probability Probability
90% [0.861, 1.747] [0.2082] [0.0529]
95% [0.809, 1.862]
99% [0.721, 2.098]
SQUARED RETURN Hurst-mandelbrot R/S Lo R/S GPH GSP
d parameter - - 0.18836 (0.0163172) 0.193619 (0.0116342)
Test Statistics 6.95016 6.1849
Critical values Probability Probability
90% [0.861, 1.747] [0.0000] [0.0000]
95% [0.809, 1.862]
99% [0.721, 2.098]

Table 4 : Long Memory Test for Silver return and Silver squared return

Hurst-mandelbrot R/S Lo R/S GPH GSP

d parameter - - 0.0092(0.0163172) 0.002(0.0116342)
Test Statistics 1.4115 1.4015
Critical values Probability Probability
90% [0.861, 1.747] [0.5696] [0.859]
95% [0.809, 1.862]
99% [0.721, 2.098]
Hurst-mandelbrot R/S Lo R/S GPH GSP
d parameter - - 0.152654 (0.0163172) 0.1494 (0.0116342)
Test Statistics 4.83328 4.4280
Critical values Probability Probability
90% [0.861, 1.747] [0.0000] [0.0000]
95% [0.809, 1.862]
99% [0.721, 2.098]

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 21
 Table 5: Long Memory Test for Crude oil return and Crude oil squared return

Hurst-mandelbrot R/S Lo R/S GPH GSP

d parameter - - -0.0085 (0.0163172) -0.0220 (0.0116342)
Test Statistics 1.3682 1.402
Critical values Probability Probability
99% [0.861, 1.747] [0.6002] [0.0585]
95% [0.809, 1.862]
99% [0.721, 2.098]
Hurst-mandelbrot R/S Lo R/S GPH GSP
d parameter - - 0.16218 (0.0163172) 0.181869 (0.0116342)
Test Statistics 4.8752 4.392
Critical values Probability Probability
90% [0.861, 1.747] [0.0000] [0.0000]
95% [0.809, 1.862]
99% [0.721, 2.098]

Table 6: Long Memory Test for Palladium return and Palladium squared return

  Hurst-mandelbrot R/S Lo R/S GPH GSP

d parameter - - 0.01508 (0.0163172) -0.0048 (0.0116342)
Test Statistics 1.352 1.348
Critical values Probability Probability
90% [0.861, 1.747] [0.3551] [0.6790]
95% [0.809, 1.862]
99% [0.721, 2.098]
  Hurst-mandelbrot R/S Lo R/S GPH GSP
d parameter - - 0.141 (0.0163172) 0.139 (0.0116342)
Test Statistics 3.654 3.391
Critical values Probability Probability
90% [0.861, 1.747] [0.0000] [0.0000]
95% [0.809, 1.862]
99% [0.721, 2.098]

22 Pınar KAYA, Bülent GÜLOĞLU

Table 7: Long Memory Test for Platinium return and Platinium squared return

  Hurst-mandelbrot R/S Lo R/S GPH GSP

d parameter - - 0.0114 (0.0163172) 0.00483 (0.0116342)
Test Statistics 1.605 1.598
Critical values Probability Probability
90% [0.861, 1.747] [0.4832] [0.6780]
95% [0.809, 1.862]
99% [0.721, 2.098]
  Hurst-mandelbrot R/S Lo R/S GPH GSP
d parameter - - 0.204 (0.0163172) 0.171 (0.0116342)
Test Statistics 4.586 4.190
Critical values Probability Probability
90% [0.861, 1.747] [0.0000] [0.0000]
95% [0.809, 1.862]
99% [0.721, 2.098]

4.3. GARCH models

The GARCH (Bollerslev, 1986) models assume that the market variance is based
on both past conditional market variance and past market shocks. Generalized Auto
Regressive Conditional Heteroskedasticity GARCH (p, q) process is given by:

α0 + ∑i 1=
βi X i + ∑ j 1ψ j Rt − j + ε t (8)
k h
Rt =
=

where ε t Ωt −1 � N (0, ht )

ω + ∑i 1=
βi ht −i + ∑ j 1α j ε t2− j (9)
p q
h=
t =

ht σ t2 Ωt −1
where =
 p q

0,  ∑ α i + ∑ β i < 1 . ε
The parameters in this model should satisfy ω > 0, α > 0, β ≥=  i 1 =i 1  t
represents disturbance term for the mean equation, Rt denotes the return of the
asset at time t, and X ’s are explanatory variables.

Equation (8) is the mean equation while Equation (9) is the conditional variance
equation. Note that in the GARCH model the parameters are restricted to be strictly
non-negative in order to satisfy the positive variance condition. Therefore GARCH
models give information about the magnitude of the shock but not about its sign.

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 23
 4.3.1.Fractionally Integrated Asymmetric Power ARCH (FIAPARCH)
Model
Tse (1998) has proposed the fractionally integrated asymmetric power (FIA-
PARCH) model, which allows for long memory and asymmetry features in the con-
ditional variance. The FIAPARCH model can be written as follows:

htδ /2= ω (1 − β L ) + 1 − (1 − β L ) (1 − λ L )(1 − L )  ( ε t − γε t )

−1 −1 d δ
(11)
 
where ω 0, δ > 0, β  1, and λ < 1 .

In the above equation (11) if the fractional integration parameter d is between
zero and 1 (0 ≤ d ≤ 1), then volatility displays the long memory property. The asy-
mmetric parameter γ satisfies the condition −1 < γ < 1. If γ > 0 it means that nega-
tive shocks are more effective on volatility than positive shocks of equal dimension.
When γ = 0 and δ = 2, the FIAPARCH model reduces to the FIGARCH model, and
when d = 0 it degrades to the APARCH model.

Conrad and Haag (2006) have created necessary and sufficient conditions for the
positivity of the conditional variance in the FIGARCH model. According to Conrad
et al. (2011), non-negativity condition for the conditional variance ht is sufficient for
all t when γ > −1 and the parameter combination (λ, d, β) satisfies the inequality
constraints, which are:

(i) even if all parameters are nonnegative, the conditional variance can become
negative and

(ii) even if all parameters are negative (apart from d ), the conditional variance
can be nonnegative almost surely.

4.3.2. Hyperbolic GARCH (HYGARCH) Model

Hyperbolic GARCH is derived by Davidson (2004). The hyperbolic GARCH
(HYGARCH) model extends the conditional variance of the FIGARCH model by int-
roducing weights into the difference operator. The HYGARCH model allows for mo-
deling long memory property in conditional volatility with hyperbolic convergence
rates. The HYGARCH (1,d,1) model can be written as follows:

 { ( d
 )}
ht = ω + 1 − (1 − β L ) λ L 1 + α (1 − L ) − 1  ε t2
−1
(12)

where ω > 0, α ≥ 0, β ≺ 1, λ < 1, and 0 ≤ d ≤ 1.

24 Pınar KAYA, Bülent GÜLOĞLU

 Davidson (2004) argues that the HYGARCH model allows for the existence of the
second moment and greater extremes of amplitudes and it could be considered as
a more general version of FIGARCH.

In fact, the hyperbolic GARCH model can be considered as a more general ver-
sion of the FIGARCH model with hyperbolic convergence rates, and permits even
more extreme amplitudes than the simple IGARCH and FIGARCH models.

4.4.Choosing between models

4.4.1. Forecast evaluation
Measuring comparative performance of out of sample forecasts obtained from
various GARCH type models is an essential part of an empirical study. In this paper
among the various forecast evaluation criteria, Mean Absolute Error (MAE), and
Theil Inequality Coefficient (TIC) are used to measure the out-of-sample forecasting
performance of the various GARCH-class models.

We use the rolling forecasting methodology in order to constitute the one- and
twenty-day out-of-sample forecasts of the various volatility models like previous stu-
dies (Chkili et al., Arouri et al., 2012b; Kang et al., 2009). The best forecasting
GARCH-based model for forecasting the volatility of commodity returns is the one
that creates the lowest prediction error.

4.4.2. The VaR and backtesting

Value at Risk is a measure that determines the potential loss in value of a risky
asset or portfolio over a certain period of time for a given confidence level. The sig-
nificance level α (confidence level 1- α ) and the risk horizon (h), which is the period
of time in terms of trading days, constitute two main parameters of VaR. According
to the Basel II Accord, banks should measure VaR at the 99 % confidence level and
use internal VaR models in order to determine their market risk capital requirement.

Like other financial assets, commodities are sensitive to market-oriented fluctua-

tions. Because of this, investors who hold a portfolio of commodities are interested
in measuring market risk of their portfolio (Chikili et al, 2014). VaR is an attractive
risk metric because it measures not only the risk factors but also the sensitivity to
risk factor. One of the most important characteristics of VaR is its universality as ap-
plicable to all activities and to all type of risk (Carol, 2008).

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 25
 While an attractive risk metric, the Value-at-Risk has however some drawbacks.
One of these is that it is not a coherent measure of risk from the viewpoint of Artz-
ner et al. (1999). A coherent risk measure should satisfy the four axioms of transla-
tion invariance, subadditivity, positive homogeneity, and monotonicity (Artzner et
al. 1999). VaR can identify extreme events, however it need not be sub-additive
which means the total risk on a portfolio should never exceed the sum of individual
risks. In order to deal with this shortcoming “Expected Shortfall” is used as a mea-
sure of risk (Scaillet, 2000).

Expected shortfall is a coherent measure for such risk and it is used to predict
the expected value of the losses conditional on the loss larger than the VaR (Scaillet,
2004). ESF can be defined as follows:

ESFt = E(|Lt | > |VaRt|), where Lt is the expected value of loss if a VaRt violation
occurs.

In this study, both long and short trading positions are taken into account and
the VaRs are estimated for the GARCH, IGARCH, FIAPARCH, and FIAPARCH models
under skewed student distribution. The daily VaR for long and short trading posi-
tions at time t can be calculated as

VaRL=
,t µˆ t + zα σˆ t and VaRS=
,t µˆ t + z1−α σˆ t
Where zα represents the left quantile at α percent of the normal distribution
and z1−α is the right quantile at α percent. µˆ t denotes to the estimated daily
conditional mean whereas σˆ t represents the estimated standard deviation of the
commodity returns obtained from a GARCH-class model.

The daily VaR’s for the skewed-Student distribution for long and short positions
is given by

VaRL=
,t µˆ t + skstα σˆ t and VaRS=
,t µˆ t + skstα σˆ t
VaR estimations should be backtested for their reliability and consistency. This
testing method enables us to compare actual profits and losses with projected VaR.
The most widely known test is Kupiec’s (1995) POF-test that examines the frequency
of losses in excess of VaR. This test defined as a likelihood ratio test (LR) which ex-
amines whether the failure rate of the model is statistically equal to the expected
one.

26 Pınar KAYA, Bülent GÜLOĞLU

 N = ∑ t =1 I t is the number of exceptions in the sample size T. Then
T

1 if rt +1 < VaR t +1 t (α )

I t +1 = 
0 if rt +1 ≥ VaR t +1 t (α )

follows a binomial distribution, N~B(T,α). If p = E ( N / T ) is the expected ex-

ception frequency, then the hypothesis for testing whether the failure rate of the
model is equal to the expected one is expressed as follows: H 0 : α = α 0 and is α 0
the prescribed VaR level.

Under the null hypothesis, the Kupiec’s (1995) likelihood ratio test is given by
 N  N   N  T − N 
LRuc = N
{
−2 log α 0 (1 − α 0 )T −N
}
+ 2 log   1 −    
 T    T   
and is asymptotically χ (1) chi squared distributed with one degree of free-
2

dom. Therefore, the model will be favored for VaR prediction which exhibits the
property that the unconditional coverage measured by p=E(N/T) equals the desired
coverage level p0.

5.Empirical Findings
5.1. Estimates of GARCH-type models
In this section, the findings obtained from different GARCH models are presen-
ted. Since normal distribution hypothesis is rejected for all returns, student t distri-
bution is used to model volatility.

GARCH estimation results for gold are illustrated in Table 8. As seen, for all mo-
dels, d-FIGARCH coefficients are statistically significant at 1% level, which implies
the existence of long memory in gold return.

The conditional volatility of gold returns reacts asymmetrically to shocks since

the APARCH (γ) parameter is negative and significant. This implies that positive and
negative shocks have an asymmetric impact on conditional volatility. Positive shocks
have an impact on the conditional volatility more than negative shocks. Among all
models with different estimation technique and distributions, FIAPARCH (1, 0.894,
1) appears to be the best model according to AIC.

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 27
 Table 8: Estimation results of GARCH methods for gold

GARC IGARC FIGARC FIAPARC FIAPARC HYGARC

 
H H H H H H
Estimation Method    
    skewed  
Cst(M) 0.058*** 0.058 0.055*** 0.061*** 0.052*** 0.056***
Cst(V) 0.011*** 0.007*** 0.012*** 0.017*** 0.016*** 0.018***
d-Figarch 0.903*** 0.894*** 0.89*** 0.923***
ARCH(Alpha1) 0.037*** 0.041*** 0.028*** 0.015 0.014 0.017
GARCH(Beta1) 0.955*** 0.958*** 0.929*** 0.925*** 0.92*** 0.928***
APARCH(Gamma1) -0.284*** -0.28
APARCH(Delta) 1.556*** 1.54***
Student(DF) 4.769*** 4.372*** 4.359*** 4.874*** 4.745***
Asymmetry -.028
Tail 4.94***
Log Alpha (HY) -0.011*
No. Observations 3693 3693 3693 3693 3693 3693
No. Parameters 5 4 6 8 9 7
Log Likelihood -5348.61 -5350.9 -5346.43 -5335.9 -5335 -5344.2
AIC 2.899 2.900 2.898 2.894 2.894 2.898
SW 2.907 2.906 2.908 2.907 2.909 2.909
SB 2.899 2.900 2.898 2.894 2.894 2.898
H-quinn 2.902 2.902 2.902 2.898 2.899 2.902
JB 4520*** 5287*** 8672.9 12759 12986 7010.9
Nyblom stability test 2.032 1.894 2.157 2.64 2.86 2.3
Pearson (50) 387.9*** 359.1*** 387.8*** 370.78*** 409.42*** 378.87***

Note: *, ** and *** indicate that statistics are significance at the 10%, 5% and 1% level of significant respectively.

In Table 9, the estimated GARCH models for copper returns are compared. Si-
milarly for all models, d-FIGARCH coefficients are statistically significant at 1% level
implying the existence of long memory in copper return.

ARCH and GARCH coefficients are also highly significant for all models except
IGARCH model. GARCH coefficient is generally greater than 0.9 which implies highly
persistence nature of volatility. GARCH coefficients of the FIGARCH, FIAPARCH, and
HYGARCH models decrease gradually because these models take into consideration
of long memory property in return series.

In addition, the conditional volatility of copper return reacts asymmetrically to

shocks since the APARCH (γ) parameter is positive and significant. Therefore, nega-
tive shocks have impacts on volatility more than positive shocks. The APARCH (δ)
power coefficient is also positive and significant.

28 Pınar KAYA, Bülent GÜLOĞLU

 Student t distribution reveals better performance than the normal distribution
and the skewed student-t distribution since the t-statistics of the Student DF para-
meter is significant at 1% level for all estimation methods.

According to AIC, FIAPARCH (1, 0.419, 1) can be chosen as to be the most app-
ropriate model.

Table 9: Estimation results of GARCH methods for copper

GARC IGARC FIGARC FIAPARC FIAPARC HYGARC

 
H H H H H H
Estimation Method    
    Skewed  
Cst(M) 0.036*** 0.035* 0.0358* 0.029 0.02 0.036*
AR(1) -0.069*** -0.069*** -0.067*** -0.067*** -0.06***
Cst(V) 0.014*** 0.01*** 0.05**
d-Figarch 0.492*** 0.419*** 0.49*** 0.439***
ARCH(Alpha1) 0.05*** 0.052*** 0.262*** 0.251*** 0.22*** 0.268***
GARCH(Beta1) 0.945*** 0.947 0.717*** 0.635*** 0.68*** 0.686***
APARCH(Gamma1) 0.124** 0.13**
APARCH(Delta) 2.137*** 1.93***
Student(DF) 6.863*** 6.528*** 7.648*** 7.008*** 6.78***
Asymmetry -.015
Tail 7.44***
Log Alpha (HY) 0.04*
No. Observations 3693 3693 3693 3693 3693 3693
No. Parameters 6 5 6 8 9 7
Log Likelihood -6781.97 -6782.71 -6786.99 -6780.13 -6788.6 -6783.2
AIC 3.676 3.675 3.678 3.676 3.681 3.677
SW 3.686 3.684 3.688 3.689 3.690 3.689
SB 3.676 3.675 3.678 3.676 3.681 3.677
H-quinn 3.679 3.678 3.682 3.681 3.686 3.681
JB 677.9 693.4 709.8 680.7 625.8 690.61
Nyblom stability test 2.518 2.29 2.54 4.946 4.5 2.855
Pearson (50) 100.29*** 97.29*** 104.71*** 104.92*** 260.8*** 97.45***

Note: *, ** and *** indicate that statistics are significance at the 10%, 5% and 1% level of significant respectively.

The results for silver and crude oil are illustrated in Table 10 and 11. The best mo-
del for silver is FIAPARCH (1, 0.414, 1) while it is FIAPARCH (1, 0.46, 1) with skewed
student distribution for crude oil. The conditional volatility of silver returns reacts
asymmetrically to shocks since the APARCH (γ) parameter is negative and signifi-
cant implying that positive shocks have more impacts on conditional volatility than

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 29
 negative shocks. The power coefficient APARCH (δ) is also positive and significant.
On the other hand, asymmetric volatility parameter (γ) is positive and significant
for crude oil volatility estimation in Table 11. This result reveals that past negative
shocks are more effective than past positive shocks on current conditional volatility.

Table 10: Estimation results of GARCH methods for silver

GARC IGARC FIGARC FIAPARC FIAPARC HYGARC

 
H H H H H H
Estimation Method         skewed  
Cst(M) 0.086*** 0.086*** 0.081*** 0.092*** 0.057** 0.085***
Cst(V) 0.024** 0.022*** 0.05
d-Figarch 0.452*** 0.414*** 0.47*** 0.345***
ARCH(Alpha1) 0.04*** 0.041*** 0.416*** 0.408*** 0.37*** 0.478***
GARCH(Beta1) 0.958*** 0.958 0.777*** 0.742*** 0.77*** 0.761***
APARCH(Gamma1) -0.21*** -0.2***
APARCH(Delta) 2.341*** 2.17***
Student(DF) 3.774*** 3.696*** 4.672*** 3.934*** 3.818***
Asymmetry -.05***
Tail 3.98***
Log Alpha (HY) 0.11**
No. Observations 3693 3693 3693 3693 3693 3693
No. Parameters 5 4 5 7 9 6
Log Likelihood -7321 -7321.1 -7327.9 -7312.9 -7309.04 -7318.72
AIC 3.967 3.967 3.971 3.964 3.96 3.966
SW 3.975 3.973 3.979 3.975 3.97 3.976
SB 3.967 3.967 3.971 3.964 3.963 3.966
H-quinn 3.970 3.969 3.974 3.968 3.96 3.970
JB 3855.1 3898.6 3118.4 3555.7 3452.7 3009.3
Nyblom stability test 2.483 2.370 2.184 3.26 4.36 2.67
Pearson (50) 386.05*** 384.59*** 358.7*** 360.98*** 558.48*** 348.33***

Note: *, ** and *** indicate that statistics are significance at the 10%, 5% and 1% level of significant respectively.

30 Pınar KAYA, Bülent GÜLOĞLU

Table 11: Estimation results of GARCH methods for crude oil

GARC IGARC FIGARC FIAPARC FIAPARC HYGARC

 
H H H H H H
Estimation Method         skewed  
Cst(M) 0.06** 0.06** 0.066** 0.058* 0.02 0.066***
AR(1) -0.046*** -0.046*** -0.04*** -0.048*** -0.047***
Cst(V) 0.024** 0.018*** 0.076*
d-Figarch 0.517*** 0.416*** 0.46*** 0.459***
ARCH(Alpha1) 0.053*** 0.055*** 0.359*** 0.406*** 0.37*** 0.384***
GARCH(Beta1) 0.943*** 0.944 0.774*** 0.718*** 0.72*** 0.752***
APARCH(Gamma1) 0.296*** 0.34***
APARCH(Delta) 1.98*** 1.76***
Student(DF) 7.064*** 6.834*** 7.995*** 7.433*** 7.06***
Asymmetry -0.05***
Tail 7.62***
Log Alpha (HY) 0.04**
No. Observations 3693 3693 3693 3693 3693 3693
No. Parameters 6 5 6 8 9 7
Log Likelihood -7863.4 -7863.85 -7867.6 -7852.6 -7854.6 -7862.8
AIC 4.261 4.261 4.264 4.257 4.25 4.262
SW 4.271 4.269 4.274 4.270 4.27 4.273
SB 4.261 4.261 4.264 4.257 4.25 4.262
H-quinn 4.265 4.264 4.267 4.261 4.26 4.266
JB 933.05 940.73 1000.3 942.27 901.5 999.83
Nyblom stability test 2.457 1.712 2.18 2.94 3.55 2.68
Pearson (50) 128.27*** 132.68*** 123.9*** 124.1*** 235.1*** 129.29***

Note: *, ** and *** indicate that statistics are significance at the 10%, 5% and 1% level of significant respectively.

GARCH estimation results for palladium is illustrated in Table 12. As seen for all
models, d-FIGARCH coefficients are statistically significant at 1% level implying the
existence of long memory in palladium return. On the other hand, the APARCH (γ)
parameter is positive and insignificant. Here, neither positive nor negative shocks
have the asymmetric effect on conditional volatility. The APARCH (δ) power coeffi-
cient is also positive and significant. According to AIC, FIAPARCH model could be
chosen to be the most appropriate model. Regarding the log likelihood, FIAPARCH
model with skewed student-t distributed innovations presents the most accurate
model.

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 31
 Table 12: Estimation results of GARCH methods for Palladium

GARC IGARC FIGARC FIAPARC FIAPARC HYGARC

 
H H H H H H
Estimation Method         skewed  
Cst(M) 0.022 0.022 0.025 0.023 0.004 0.026
Cst(V) 0.059*** 0.058*** 0.109*** 0.109*** 0.109*** 0.078
d-Figarch 0.578*** 0.571*** 0.571*** 0.53***
ARCH(Alpha1) 0.10*** 0.10*** 0.27*** 0.289*** 0.288*** 0.297***
GARCH(Beta1) 0.899*** 0.899*** 0.709*** 0.708*** 0.708*** 0.694***
APARCH(Gamma1) 0.043 0.044
APARCH(Delta) 1.98*** 1.98***
Student(DF) 3.985*** 3.97*** 4.19*** 4.19*** 4.036***
Asymmetry -.031*
Tail 4.19***
Log Alpha (HY) 0.035
No. Observations 3693 3693 3693 3693 3693 3693
No. Parameters 5 4 6 8 9 7
Log Likelihood -7383.2 -7383.2 -7376.6 -7376.6 -7375 -7376.1
AIC 4 4 4 3.998 3.998 3.998
SW 4 4 4.008 4.012 4.014 4.010
SB 4 4 3.998 3.998 3.998 3.998
H-quinn 4 4 4 4.003 4 4.002
JB 16344 16576 15382 14914 14909 18114
Nyblom stability test 2.241 2 3 4.479 4.94 3.44
Pearson (50) 1704.25*** 1703.8*** 1713.29*** 1717.04*** 1709.55*** 1706.95***

Note: *, ** and *** indicate that statistics are significance at the 10%, 5% and 1% level of significant respectively.

In Table 13, various GARCH estimation results for platinum return are displayed.
The results show evidence of long memory in platinum return. The APARCH (δ) po-
wer coefficient is also positive and significant. However, the APARCH (γ) parameter
is negative and insignificant. Here it couldn’t be mentioned the asymmetric effects
of positive and negative shocks. According to information criteria and log likelihood
ratio, FIAPARCH (1,0.74,1) model with skewed student-t distributed innovations
could be interpreted as the most appropriate model.

32 Pınar KAYA, Bülent GÜLOĞLU

Table 13: Estimation results of GARCH methods for Platinum

GARC IGARC FIGARC FIAPARC FIAPARC HYGARC

 
H H H H H H
Estimation Method         skewed  
Cst(M) 0.046*** 0.046*** 0.045*** 0.052*** 0.038** 0.03
Cst(V) 0.029*** 0.019*** 0.049** 0.049 0.051 0.06
d-Figarch 0.592*** 0.767*** 0.746*** 0.764***
ARCH(Alpha1) 0.07*** 0.07*** 0.274*** 0.177 0.185 0.177
GARCH(Beta1) 0.916*** 0.922*** 0.75*** 0.841*** 0.832*** 0.816***
APARCH(Gamma1) -0.039 -0.042
APARCH(Delta) 1.432*** 1.469***
Student(DF) 5.16*** 4.72*** 4.9*** 5.27***
Asymmetry -.037 -0.043**
Tail 5.3*** 5.22***
Log Alpha (HY) -0.032
No. Observations 3693 3693 3693 3693 3693 3693
No. Parameters 5 4 6 8 9 8
Log Likelihood -5938.4 -5941.06 -5938.57 -5931.6 -5930.14 -5.935
AIC 3.218 3.219 3.219 3.216 3.216 3.218
SW 3.227 3.219 3.229 3.230 3.231 3.231
SB 3.218 3.226 3.219 3.216 3.216 3.218
H-quinn 3.221 3.222 3.222 3.221 3.221 3.223
JB 867.96 898.6 857.75 855.49 859.55 858.97
Nyblom stability test 3.728 3.368 4.505 5 5.31 5.09
Pearson (50) 200.54*** 211.1*** 206.6*** 252.85*** 443.59*** 327.21***

Note: *, ** and *** indicate that statistics are significance at the 10%, 5% and 1% level of significant respectively.

We estimate FIAPARCH model both assuming Student-t and skewed Student-t

distributed innovations. Asymmetric parameters are negative and statistically signi-
ficant for silver, crude oil and palladium respectively at the %1, %5 and %10 level.
Therefore, one may observe from these results that silver, crude oil and palladium
innovations are skewed to the left. In addition, the tail parameters in all the FIA-
PARCH models are statistically significant and positive. This reveals that the commo-
dity return series are fat-tailed. To sum up, FIAPARCH model with skewed student-t
distributed innovations is the most accurate model for most of these commodity
return series. It takes into consideration both asymmetry and long memory proper-
ties of series.

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 33
 5.2. Forecast evaluation
To evaluate the forecast performance of different GARCH models, we constitute
the one and twenty-day-ahead volatility forecasts over the out of sample period
from February 29, 2016 through April 29, 2016 to assess the forecasting perfor-
mance of the eight competing GARCH models. Theil Inequality Coefficient (TIC) and
Mean Absolute Error (MAE) are reported in tables 14, 15 for the mean equation of
these commodities. As can be seen from Table 14, with regard to the one-day fore-
casting horizon, FIAPARCH model with student distribution outperforms the other
models for metals but HYGARCH model performs better for crude oil, palladium
and platinum.

Table 14: Forecast Comparison of Volatility models

Model Criteria GOLD COPPER SILVER CRUDE OIL PALLADIUM PLATINIUM

GARCH MAE 0.623 0.077 0.655 3.36 0.383 1.021

TIC 0.84 0.247 0.79 0.984 0.896 1

IGARCH MAE 0.624 0.076 0.655 3.36 0.383 1.021

TIC 0.842 0.242 0.791 0.984 0.897 1

FIGARCH MAE 0.626 0.081 0.659 3.354 0.379 1.02

TIC 0.849 0.26 0.801 0.981 0.879 1

FIAPARCH MAE 0.62 0.075 0.649 3.363 0.382 1.027

TIC 0.833 0.236 0.779 0.986 0.891 1

FIAPARCH MAE 0.63 0.21 0.68 3.35 0.401 1.014

skewed TIC 0.85 1 0.85 0.98 0.978 1

HYGARCH MAE 0.626 0.079 0.656 3.354 0.379 1.005

  TIC 0.847 0.255 0.792 0.981 0.879 1

Notes: This table reports the results of the one-day out-of-sample prediction errors of return series for the different

competing volatility models.

When looking at the results of twenty-day forecast horizon in Table 15, it can be
seen that the FIAPARCH model has superiority in terms of volatility forecast accuracy
for all commodity returns except that of copper and palladium. With respect to Theil
Inequality Coefficient (TIC), while IGARCH model produces the lowest mean loss for

34 Pınar KAYA, Bülent GÜLOĞLU

copper, FIGARCH and HYGARCH model produce that of for Palladium and the latter
produces the lowest mean absolute error for Platinum.

Table 15: Forecast Comparison of Volatility models

Model Criteria GOLD COPPER SILVER CRUDE OIL PALLADIUM PLATINIUM

GARCH MAE 0.989 1.043 1.189 2.849 1.63 1.396

TIC 0.959 0.965 0.948 0.979 0.986 0.972
IGARCH MAE 0.989 1.043 1.189 2.849 1.63 1.396
TIC 0.959 0.965 0.948 0.979 0.986 0.972
FIGARCH MAE 0.988 1.043 1.188 2.847 1.629 1.396
TIC 0.961 0.966 0.95 0.977 0.983 0.973
FIAPARCH MAE 0.989 1.043 1.19 2.84 1.63 1.397
TIC 0.957 0.969 0.945 0.98 0.985 0.969
FIAPARCH MAE 0.98 1.05 1.18 2.85 1.637 1.396
skewed TIC 0.96 0.98 0.96 0.99 0.997 0.977
HYGARCH MAE 0.988 1.043 1.189 2.847 1.629 1.395
  TIC 0.96 0.966 0.948 0.977 0.983 0.981

Notes: This table reports the results of the Twenty-day out-of-sample prediction errors of return series
for the different competing volatility models.

Once again, out of sample forecast analysis implies that allowing for asymmetry
and long memory properties leads to enhance the quality of volatility forecasts of
commodity returns. In most cases, volatility estimates show that the FIAPARCH mo-
del is favored to the other five GARCH-class models.

5.3 VaR estimations

In this part of our analysis, VaR of commodities are estimated based on three
GARCH models for level α from 5% to 0.25%. Besides we compute the failure rate
for the long trading position as a percentage of negative returns smaller than one-
step-ahead VaR for long positions. For the short trading position, the failure rate is
defined as the percentage of positive returns larger than the one-step-ahead VaR
for short positions.

Table 16 shows the in sample VaR estimations based on GARCH, FIGARCH, and
FIAPARCH models. In this respect, the Kupiec LR tests are carried out for each of
these GARCH models, which investigate whether the empirical failure rate is equal
to the pre-specified VaR level α . The other risk measures Expected Shortfalls (ESF)
are presented for long and short positions and labelled ESF1 and ESF2 respectively

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 35
 in Table 17-18. Hendricks (1996) defines the ESF1 as the excess value of the losses
over the VaR, the ESF2 as the expected value of loss exceeding the VaR level, divided
by the associated VaR values.

According to our results, expected shortfalls are the highest for crude oil and the
lowest for gold for all risk levels. The Kupiec test statistics are highly significant at
all levels (1%, 5%, and 10%) in most cases. This result suggests that the standard
GARCH model performs inadequately at all events regardless of the commodity
analyzed. For the short trading position, the use of the FIGARCH model for Copper,
Silver, and Crude oil returns performs well. On the other hand, for these three com-
modity return series, the Kupiec test statistics are significant at three conventional
levels for the long trading positions with the FIGARCH model. However, the Kupiec
test statistics are insignificant for all commodities and for long trading position at
% 1 level of significance with skewed t distribution FIAPARCH model. More preci-
sely the findings show that the FIAPARCH model with skewed student distributed
innovations outperforms the other methods for long position and it has more or less
similar performance with FIGARCH model for short trading positions. It can be conc-
luded that in the models which take into account volatility clustering, asymmetry
and long memory in the commodity returns, VaR and ESF are more accurate measu-
re of risk for both long and short trading positions.

36 Pınar KAYA, Bülent GÜLOĞLU

Table 16: Kupiec LR test statistics based on in sample VaR

Short positions   Long positions

FIAPARCH FIAPARCH
α quantile GARCH FIGARCH FIAPARCH α quantile GARCH FIGARCH FIAPARCH
skewed skewed
GOLD GOLD
0.95 0.2274 0.0103 0.0024 0.39 0.05 7.105*** 2.991* 7.885*** 4.682**
0.975 0.9993 3.998** 1.766 0.61 0.025 6.244*** 1.984 6.74*** 3.27*
0.99 15.214*** 15.214*** 10.694*** 10.69*** 0.01 0.9572 0.0238 3.613** 1.66
0.995 6.018*** 11.483*** 7.576*** 6.01*** 0.005 1.184 2.598 0.6984 1.81
0.9975 5.7308*** 12.038*** 8.36*** 3.78** 0.0025 0.1728 0.5908 0.59088 1.29
COPPER COPPER
0.95 0.1242 1.483 0.391 0.39 0.05 0.993 4.078** 0.6001 0.00069
0.975 0.612 1.006 0.005 0.14 0.025 1.455 4.437** 0.637 0.078
0.99 2.97* 0.023 1.856 1.85 0.01 0.957 8.588*** 2.087 0
0.995 7.576*** 2.598 6.018*** 1.18 0.005 0.011 2.091 0.011 0.011
0.9975 3.78** 3.78** 3.78** 3.78* 0.0025 1.366 3.033* 0.319 0.31
SILVER SILVER
0.95 0.78963 1.8621 2.2794 0.015 0.05 5.662*** 15.471*** 8.7031*** 3.24*
0.975 6.614*** 0.2389 4.9591** 0.019 0.025 4.862** 27.275*** 6.7408*** 2.27
0.99 10.694*** 0.24122 7.1297*** 2.97* 0.01 0.4374 8.5883*** 2.5534 0.0238
0.995 7.5768*** 1.1845 9.3862*** 3.54** 0.005 1.038 9.3092*** 1.038 0.015
0.9975 .NaN 3.7809** 12.038*** 8.36*** 0.0025 0.0622 6.521*** 0.75932 0.319
CRUDE
CRUDE OIL
OIL
0.95 2.521 0.107 0.789 0.72 0.05 4.078** 11.38*** 1.309 .015
0.975 2.762* 0.148 2.762* 0.124 0.025 1.221 8.334*** 2.277 .078
0.99 8.22*** 0.696 7.129*** 2.37 0.01 0.437 4.206** 0.673 0.000133
0.995 6.018*** 4.683** 7.576*** 3.54* 0.005 0.64 4.269*** 2.74* 0.334
0.9975 0.59 0.172 1.296 1.29 0.0025 1.366 5.237** 3.033* .759
PALLADIUM PALLADIUM
0.95 0.00069 0.227 0.49 2.067 0.05 0.063 1.146 0.993 0.227
0.975 0.211 0.321 0.612 0.005 0.025 0.483 0.637 0.637 0.03
0.99 2.974* 2.379 2.974* 1.017 0.01 1288 5.513*** 4.206** 1.665
0.995 0.698 0.12 0.0118 0.011 0.005 0.011 0.64 1.038 0.015
0.9975 2.336 0.5908 0.5908 0.172 0.0025 0.062 0.319 0.319 0.319
PLATINIUM PLATINIUM
0.95 1.832 2.777* 1.627 .015 0.05 4.078 2.991* 4.998** 2.508
0.975 2.404 5.481*** 1.485 .321 0.025 2.277 .238 2.277 .350
0.99 1.017 2.974* .696 .103 0.01 2.087 .673 2.553 .673
0.995 1.184 3.549*** 1.814 .346 0.005 .334 .015 1.038 .124
0.9975 .590 1.296 .590 .172 0.0025 .172 1.296 0 5.730***
Notes: The table reports the Kupiec test statistics. *, ** and *** indicate that statistics are significance at the 10%, 5% and
1% level of significant respectively. The best model is the one with the least rejections.

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 37
 Table 17: Expected Shortfalls based on in-sample Value-at-Risk

Short positions
FIAPARCH FIAPARCH
α quantile GARCH GARCH FIGARCH FIGARCH FIAPARCH FIAPARCH
skewed skewed
  ESF1 ESF2 ESF1 ESF2 ESF1 ESF2 ESF1 ESF2
GOLD
0.95 2.42 1.3 2.38 1.28 2.41 1.3 2.39 1.31
0.975 2.82 1.2 2.82 1.19 2.88 1.22 2.84 1.23
0.99 3.95 1.22 3.9 1.17 3.9 1.19 3.9 1.23
0.995 4.55 1.1 4.49 1.11 4.64 1.14 4.43 1.15
0.9975 5.11 1.02 3.66 1 5.84 1.07 5.32 1.06
COPPER
0.95 3.53 1.3 3.4 1.31 3.43 1.3 3.48 1.31
0.975 4.1 1.21 3.98 1.23 3.96 1.22 4.07 1.21
0.99 5.07 1.13 4.69 1.15 4.83 1.16 5.01 1.17
0.995 5.71 1.11 4.83 1.12 5.16 1.14 5.02 1.1
0.9975 7.14 1.04 7.14 1.12 7.15 1.08 7.14 1.1
SILVER
0.95 4.17 1.34 3.91 1.37 4.17 1.36 4.07 1.37
0.975 5.05 1.28 4.55 1.29 5 1.27 4.76 1.26
0.99 6.07 1.18 5.37 1.22 5.67 1.18 5.49 1.19
0.995 8 1.06 6.61 1.16 8.05 1.1 6.91 1.11
0.9975 .NaN .NaN 5.48 1.08 4.7 1.03 9.18 1.07
CRUDE OIL
0.95 4.9 1.31 4.78 1.31 4.86 1.29 4.76 1.29
0.975 5.91 1.24 5.56 1.24 5.68 1.23 5.61 1.24
0.99 7.31 1.29 6.63 1.22 7.2 1.24 6.93 1.23
0.995 8.86 1.38 8.82 1.37 9.14 1.34 8.16 1.29
0.9975 9.66 1.28 9.14 1.27 9.62 1.26 9.62 1.31
PALLADIUM
0.95 4.23 1.44 4.21 1.43 4.18 1.43 4.12 1.44
0.975 5.15 1.37 5.24 1.39 5.27 1.4 5.09 1.4
0.99 6.56 1.42 6.54 1.43 6.55 1.45 6.34 1.43
0.995 7.21 1.37 7.1 1.35 6.99 1.35 6.99 1.39
0.9975 8.24 1.58 8.67 1.45 8.67 1.47 8.06 1.45
PLATINIUM
0.95 2.87 1.38 2.81 1.38 2.9 1.39 2.82 1.39
0.975 3.47 1.33 3.48 1.34 3.44 1.33 3.37 1.34
0.99 4.16 1.28 4.17 1.29 4.26 1.28 4.12 1.29
0.995 4.98 1.29 5.27 1.31 5.14 1.31 4.79 1.29
0.9975 5.4 1.29 5.52 1.27 5.4 1.28 5.78 1.29

38 Pınar KAYA, Bülent GÜLOĞLU

Table 18: Expected Shortfalls based on in-sample Value-at-Risk

Long positions
FIAPARCH FIAPARCH
α quantile GARCH GARCH FIGARCH FIGARCH FIAPARCH FIAPARCH
skewed skewed
  ESF1 ESF2 ESF1 ESF2 ESF1 ESF2 ESF1 ESF2
GOLD
0.05 -2.46 1.45 -2.5 1.46 -2.46 1.45 -2.49 1.45
0.025 -3.01 1.35 -3.02 1.36 -2.95 1.36 -3.01 1.36
0.01 -3.83 1.29 -3.85 1.31 -3.77 1.26 -3.83 1.26
0.005 -4.72 1.41 -4.81 1.44 -4.85 1.4 -4.99 1.44
0.0025 -5.55 1.39 -5.45 1.42 -5.59 1.49 -5.48 1.54
COPPER
0.05 -3.6 1.41 -3.5 1.43 -3.54 1.4 -3.6 1.39
0.025 -4.06 1.34 -4.01 1.37 -4.1 1.34 -4.28 1.33
0.01 -4.95 1.31 -4.58 1.29 -4.64 1.28 -4.97 1.31
0.005 -5.67 1.4 -5.35 1.36 -5.5 1.39 -5.66 1.35
0.0025 -6.17 1.32 -5.87 1.37 -6.23 1.38 -6.23 1.35
SILVER
0.05 -4.54 1.55 -4.31 1.57 -4.45 1.53 -4.61 1.51
0.025 -5.63 1.44 -5.03 1.43 -5.51 1.43 -5.73 1.42
0.01 -7.34 1.43 -6.58 1.43 -6.9 1.39 -7.45 1.43
0.005 -8.25 1.35 -7.54 1.38 -8.15 1.38 -8.74 1.38
0.0025 -9.97 1.39 -8.24 1.35 -9.51 1.33 -9.9 1.29
CRUDE OIL
0.05 -4.6 1.38 -4.54 1.41 -4.71 1.41 -4.86 1.4
0.025 -5.39 1.34 -5.3 1.34 -5.67 1.34 -5.7 1.34
0.01 -6.01 1.31 -5.53 1.34 -6.05 1.34 -6.52 1.33
0.005 -6.93 1.3 -6.31 1.34 -6.56 1.3 -7.08 1.31
0.0025 -8.15 1.28 -7.24 1.33 -7.64 1.3 -8.27 1.31
PALLADIUM
0.05 -4.64 1.51 -4.56 1.49 -4.57 1.49 -4.64 1.48
0.025 -5.54 1.43 -5.74 1.44 -5.74 1.43 -5.9 1.43
0.01 -6.95 1.31 -6.56 1.29 -6.69 1.29 -6.87 1.29
0.005 -8.95 1.3 -7.95 1.26 -7.89 1.25 -8.22 1.26
0.0025 -9.67 1.21 -9.88 1.22 -9.88 1.21 -9.88 1.18
PLATINIUM
0.05 -3 1.4 -3 1.39 -2.99 1.41 -3.05 1.39
0.025 -3.73 1.33 -3.7 1.33 -3.7 1.33 -3.88 1.33
0.01 -4.57 1.23 -4.61 1.21 -4.64 1.24 -4.67 1.23
0.005 -5.39 1.18 -5.48 1.15 -5.36 1.17 -5.6 1.16
0.0025 -6.42 1.14 -6.37 1.14 -6.42 1.15 -9.05 1.35

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 39
 5.4 Out of Sample Forecast Results
As mentioned above, the FIAPARCH model with skewed t seems to outperforms
the others in terms of forecasting performance. Moreover, 95 % the value at risk
obtained from that model passes the backtesting (Kupiec test) and confirms that
value at risk captures well the losses and gains for short and long positions respec-
tively. Based on these considerations we perform out of sample forecasts for con-
ditional mean, conditional variance and value at risk. The findings are illustrated in
figures 4-9. As seen there are major differences in volatility and VaR among markets.
It seems that gold and platinum have similar volatility and VaR patterns. In both
markets, volatility diminishes over the forecast period and risk of losses decreases. In
contrast, the volatility of silver copper and palladium tends to increase rapidly over
time. As a result, in those markets value at risk for short positions tend to increase.
With respect to oil we have a somewhat different picture. As shown in figure 7 oil
volatility is much higher than the others. Moreover, the dynamic value at risk plot
also confirms that oil market risk is considerably large compared with other com-
modities.

Figure 4: Skewed-student FIAPARCH forecasting for Gold

40 Pınar KAYA, Bülent GÜLOĞLU

Figure 5: Skewed-student FIAPARCH forecasting for Copper

Figure 6: Skewed-student FIAPARCH forecasting for Silver

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 41
 Figure 7: Skewed-student FIAPARCH forecasting for Crude oil

Figure 8: Skewed-student FIAPARCH forecasting for Palladium

42 Pınar KAYA, Bülent GÜLOĞLU

Figure 9: Skewed-student FIAPARCH forecasting for Platinum

6. Conclusion
Risk measurement of an asset is essential for investors in portfolio allocation, hed-
ging and risk management. In this paper using various GARCH models, the riskiness
of gold, silver, copper, crude oil, palladium, and platinum are modelled and forecas-
ted. Volatility, value at risk, and expected shortfall are employed as risk measures.
All risk measures are comparable as they are all based on the same GARCH model.

It turns out that among existing GARCH models those allowing for long memory
and asymmetry outperform the others in terms of out of sample forecast. Moreover,
it seems that those models perform better in backtesting. In other words, the value
at risk and expected shortfall risk measures are more credible. More specifically for
most of these commodities, FIAPARCH model appears to be the best model not only
in terms of goodness of fit measures such as AIC, SW, SB and H-Quinn criteria, and
Log likelihood but also in terms of predicting volatility and value at risk. It seems that
taking into account asymmetry and long memory also improves the performance
of out of sample forecasts. In this respect, the FIAPARCH model outperforms in
most of the cases except for palladium and platinum. Other studies dealing with

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 43
 commodity markets also attain almost identical results. For instance Aloui and Mab-
rouk (2010) found that considering for long-range memory, fat-tails and asymmetry
performs better in predicting a one-day-ahead VaR for both short and long trading
positions. They also argue that the FIAPARCH model outperforms the other models
in the VaR’s prediction. Chkili et al. (2014) suggest also that the FIAPARCH model
is the best suited for estimating the VaR forecasts for both short and long trading
positions of commodities’ returns.

The results based on FIAPARCH model(with skewed t) suggest that oil volatility
is high and the risk of oil measured by value at risk and expected shortfall is signifi-
cantly higher than the other commodities considered in the study. The finding is in
contradiction to the findings obtained in previous studies. However, it is consistent
with the reality. It is well known that oil price is very sensitive to political issues.
Over recent years many political events such as the invasion of Iraq, sanctions on oil
exporting countries like Iran, Russia, Venezuela and the slowdown of Chinese eco-
nomic growth cause oil price to fluctuate more. Moreover, after the improvement
in shell gas technology, the USA becomes an oil exporting country affecting supply
and the price of oil.

The above considerations suggest that oil is no longer safe heaven and cast do-
ubt on the use of oil as a hedging tool. The results on other commodities are more
and less in accord with those obtained in earlier studies.

44 Pınar KAYA, Bülent GÜLOĞLU

 References
1. Aloui, C. and Mabrouk S.. (2010). Value-at-risk estimations of energy com-
modities via long- memory, asymmetry and fat-tailed GARCH models, Energy
Policy, 38 (5). 2326–2339.

2. Arouri, M. Hammoudeh, S. Lahiani, A. Nguyen, D. K.. (2012a). Long memory

and structural breaks in modeling the return and volatility dynamics of pre-
cious metals, The Quarterly Review of Economics and Finance, 52: 207–218.

3. Arouri, M. Lahiani, A,. Lévy, A. Nguyen, D. K.. (2012b). Forecasting the con-
ditional volatility of oil spot and futures prices with structural breaks and long
memory models, Energy Economics, 34 (1):283–293.

4. Baillie, R. T.. (1996). Long memory processes and fractional integration in

econometrics, Journal of Econometrics, 73: 5–59.

5. Baillie, R. T. Bollerslev, T. and Mikkelsen, H. O.. (1996). „Fractionally inte-

grated generalized autoregressive conditional heteroskedasticity, J. Econo-
metrics, 74:3-30.

6. Bollerslev, T.. (1986). Generalized autoregressive conditional heteroscedas-

ticity, Journal of econometrics, 31(3), 307-327.

7. Bollerslev, T. and Mikkelsen, H. O.. (1996). Modelling and pricing long mem-
ory in stock market volatility, Journal of Econometrics, 73:151–184.

8. Beine, M. and Laurent, S.. (2003). Central bank interventions and jumps in
double long memory models of daily exchange rates, Journal of Empirical
Finance, 10:641–660.

9. Cheng, W. H. and Hung, J. C.. (2011). „Skewness and leptokurtosis in

GARCH-typed VaR estimation of petroleum and metal asset returns, Journal
of Empirical Finance, 18:160–173.

10. Cheong, C. W.. (2009). Modeling and forecasting crude oil markets using
ARCH-type models, Energy Policy, 37:2346–2355.

11. Chkili, W. Hammoudeh, S. D. and Nguyen, K.. (2014). Volatility forecasting

and risk management for commodity markets in the presence of asymmetry
and long memory, Energy Economics, 41:1-18.

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 45
 12. Choi, K. and Hammoudeh, S.. (2009). Long memory in oil and refined prod-
ucts markets, Energy Journal, 30: 97–116.

13. Conrad, C.. (2010). Non-negativity conditions for the hyperbolic GARCH
model, Journal of Econometrics, 157(2):441-457.

14. Conrad, C. and Haag, B. R.. (2006). „Inequality constraints in the fractionally
integrated GARCH model, Journal of Financial Econometrics, 4:413–449.

15. Conrad, C. and Karanasos, M.. (2005a). On the inflation-uncertainty hypoth-

esis in the USA, Japan and the UK: a dual long memory approach, Japan and
the World Economy, 17:327–343.

16. Conrad, C. and Karanasos, M.. (2005b). „Dual long memory in inflation dy-
namics across countries of the Euro area and the link between inflation un-
certainty and macroeconomic performance, Studies in Nonlinear Dynamics &
Econometrics, 4 (5).

17. Conrad, C. Karanasos, M. and Zeng, N.. (2011). Multivariate fractionally in-
tegrated APARCH modelling of stock market volatility: multi-country study,
Journal of Empirical Finance, 18:147–159.

18. Creti, A. Joëts, M. and Mignon, V.. (2013). On the links between stock and
commodity markets’ volatility, Energy Econonomics, 37:16–28.

19. Dahl, C. M. and Iglesias, E. M.. (2009). „Volatility spillovers in commodity

spot prices: new empirical results, Economic Modelling, 26:601–607.

20. Davidson, J.. (2004). Moment and memory properties of linear conditional
heteroscedasticity models, and a new model, Journal of Business and Eco-
nomic Statistics, 22:16–29.

21. Ding, Z. Granger, C. W. J. and Engle, R. F.. (1993). A Long Memory Property
of Stock Market Returns and a New Model, Journal of Empirical Finance,
1:83-106 


22. Engle, R. F.. (1982). Autoregressive Conditional Heteroscedasticity with Esti-

mates of Variance of United Kingdom Inflation, Econometrica, 50 (4):987–
1008.

23. Engle, R. F. Lilien, D. M. and Robins, R. P.. (1987). Estimating time varying
risk premia in the term structure: The ARCH-M model, Econometrica: Journal

46 Pınar KAYA, Bülent GÜLOĞLU

 of the Econometric Society, 391-407.

24. Geweke, J. and Porter-Hudak, S.. (1983). The estimation and application of
long memory time series models, Journal of Time Series Analysis, 4:221-238.

25. Granger, C.. (1980). Long Memory Relationships and the Aggregation of
Dynamic Models, Journal of Econometrics, 14:227-238.

26. Granger, C. and Joyeux, R.. (1980). An Introduction to Long-Memory Time

Series Models and Fractional Differencing, Journal of Time Series Analysis,
1:15-29.

27. Hosking, J.. (1981). Fractional Differencing, Biometrika, 68:165-176.

28. Hung, J. C. Lee, M. C. and Liu, H. C.. (2008). Estimation of value-at-risk

for energy commodities via fat-tailed GARCH models, Energy Ecoonomics.
30:1173–1191.

29. Hurst, H. E.. (1951). Long-term storage capacity of reservoirs, Transactions of

the American Society of Civil Engineers, 116:770-799.

30. Jain A,.Biswal P.C(2016) Dynamic linkages among oil price, gold price, ex-
change rate, and stock market in India, Resources Policy, 49,179-185

31. Jarque, C. M. and Bera, A. K.. (1987). A Test for Normality of Observations
and Regression Residuals, International Statistical Review / Revue Internatio-
nale de Statistique, 55(2):163-172.

32. Kang, S. H. Kang, S. M. and Yoon, S. M.. (2009). Forecasting volatility of

crude oil markets, Energy Economics, 31:119–125.

33. Kang, S. H. McIver R. and Yoon, S. M. (2016). Modeling Time-Varying Cor-

relations in Volatility Between BRICS and Commodity Markets Emerging Mar-
kets Finance & Trade, 52(7), 1698-1723. 

34. Kang, S. H. McIver R. and Yoon, S. M. (2017). Dynamic spillover effects

among crude oil, precious metal, and agricultural commodity futures mar-
kets, Energy Economics, 62,19-32.

35. Kang, S. H. and Yoon, S. M.. (2013). Modeling and forecasting the volatility
of petroleum futures prices, Energy Economics, 36:354–362.

36. Kumar, D.. (2014). Long Memory in the Volatility of Indian Financial Market:

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 47
 An Empirical Analysis Based on Indian Data. Anchor Academic Publishing,
Hamburg.

37. Lo, A. W.. (1991). Long-term memory in stock market prices, Econometrica,
59:1279-1313.


38. Lobato, I. N. and Savin, N. E.. (1998). Real and spurious long memory prop-
erties of stock market data, Journal of Business and Economic Statistics,
16:261–268.

39. Mandelbrot, B. B.. (1971). When Can Price be Arbitraged Efficiently? A Limit
to the Validity of the Random Walk and Martingale Models, Review of Eco-
nomics and Statistics, 53(3):225-236.

40. Mandelbrot, B. B, and Wallis, J. R.. (1969). „Some long-run properties of

geophysical records, Water Resouces. Research, 5(2):321-340.


41. Mohammadi, H. and Su, L.. (2010). International evidence on crude oil
price dynamics: applications of ARIMA-GARCH models, Energy Economics,
32:1001–1008.

42. Muth, J. F.. (1961). Rational expectations and the theory of price move-
ments, Econometrica: Journal of the Econometric Society, 315-335.

43. Regnier, E.. (2007). Oil and energy price volatility, Energy Economics,
29:405–427.

44. Robinson, P. M.. (1991). Testing for Strong Serial Correlation and Dynamic
Conditional Heteroscedasticity in Multiple Regression, Journal of Economet-
rics, 47:67–84.

45. Robinson, P. M. and Henry, M.. (1999). Long and short memory conditional
heteroskedasticity in estimating the memory parameter of levels, Economet-
ric Theory, 15(3):299-336.

46. Thuraisamy, K. S. Sharma, S. S. and Ahmed, H. J. A.. (2013). The relation-

ship between Asian equity and commodity futures markets, Journal of Asian
Economics, 28: 67–75.

47. Tsay R S(2010). Analysis of Financial Time Series 3rd Edition, Wiley

48. Tse, Y.. (1998). The conditional heteroscedasticity of the yen–dollar ex-

48 Pınar KAYA, Bülent GÜLOĞLU

 change rate, Journal of Applied Economics, 13:49–55.

49. Vivian, A. and Wohar, M. E.. (2012). Commodity volatility breaks, Journal of
International Financial Markets Institutions and Money, 22:395–422.

50. Wang, Y. Wei, Y. and Wu, C.. (2010). Auto-correlated behavior of WTI crude
oil volatilities: A multiscale perspective, Physica A: Statistical Mechanics and
its Applications, 389(24):5759-5768.

51. Wei, Y. Wang, Y. and Huang, D.. (2010). „Forecasting crude oil market
volatility: further evidence using GARCH-class models, Energy Economics
32:1477–1484.

52. Yajima, Y.. (1985). On estimation of long memory time series models, Aus-
tralian Journal of Statistics, 27:303-20.

Modeling and Forecasting the Markets Volatility and VaR Dynamics of Commodity 49

View publication stats